Hacker News Reader

> Happy to hear that they've introduced video encoders and decoders based on compute shaders.

This is great news. I remember being laughed at when I initially asked whether the Vulkan enc/dec were generic because at the time it was all just standardising interfaces for the in-silicon acceleration.

Having these sorts of improvements available for legacy hardware is brilliant, and hopefully a first route that we can use to introduce new codecs and improve everyone's QOL.

I haven't even had a cursory look at decoders state of the art for 10+ years. But my intuition would say that decoding for display could profit a lot from GPU acceleration for later parts of the process when there is already pixel data of some sort involved. Then I imagine thet the initial decompression steps could stay on the CPU and the decompressed, but still (partially) encoded data is streamed to the GPU for the final transformation steps and application to whatever I-frames and other base images there are. Steps like applying motion vectors, iDCT... look embarrassingly parallel at a pixel level to me.

When the resulting frame is already in a GPU texture then, displaying it has fairly low overhead.

My question is: how wrong am I?

I'm not an expert, but in the worst case, you might need to decode dense 4x4-pixel blocks which each depend on fully-decoded neighbouring blocks to their west, northwest, north and northeast. This would limit you to processing `frame_height * 4` pixels in parallel, which seems bad, especially for memory-intensive work. (GPUs rely on massive parallelism to hide the latency of memory accesses.)

Motion vectors can be large (for example, 256 pixels for VP8), so you wouldn't get much extra parallelism by decoding multiple frames together.

However, even if the worst-case performance is bad, you might see good performance in the average case. For example, you might be able to decode all of a frame's inter blocks in parallel, and that might unlock better parallel processing for intra blocks. It looks like deblocking might be highly parallel. VP9, H.265 and AV1 can optionally split each frame into independently-coded tiles, although I don't know how common that is in practice.

Exciting! I am consistently blown away by the talent of the ffmpeg maintainers. This is fairly hard stuff in my opinion and they do it for free.

Could you explain more about it? I assumed the maintainers are doing it as part of their jobs for a company (completely baseless assumption)

Reupvoted you from gray because I don't think that's fair, but I also don't know how much there is to add. As far as why I'm contributing, I haven't been socially involved in the ffmpeg dev community in a decade, but, it is a very reasonable floor to assume it's 80% not full time paid contributors.

These release notes are very interesting! I spent a couple of weeks recently writing a ProRes decoder using WebGPU compute shaders, and it runs plenty fast enough (although I suspect Apple has some special hardware they make use of for their implementation). I can imagine this path also working well for the new Android APV codec, if it ever becomes popular.

The ProRes bitstream spec was given to SMPTE [1], but I never managed to find any information on ProRes RAW, so it's exciting to see software and compute implementations here. Has this been reverse-engineered by the FFMPEG wizards? At first glance of the code, it does look fairly similar to the regular ProRes.

[1] https://pub.smpte.org/doc/rdd36/20220909-pub/rdd36-2022.pdf

Do you have a link for that? I'm the guy working on the Vulkan ProRes decoder mentionned as "in review" in this changelog, as part of a GSoC project.

I'm curious wrt how a WebGPU implementation would differ from Vulkan. Here's mine if you're interested: https://github.com/averne/FFmpeg/tree/vk-proresdec

I don't have a link to hand right now, but I'll try to put one up for you this weekend. I'm very interested in your implementation - thanks, will take a good look!

Initially this was just a vehicle for me to get stuck in and learn some WebGPU, so no doubt I'm missing lots of opportunities for optimisation - but it's been fun as much as frustrating. I leaned heavily on the SMPTE specification document and the FFMPEG proresdec.c implementation to understand and debug.

No problem, just be aware there's a bunch of optimizations I haven't had time to implement yet. In particular, I'd to remove the reset kernel, fuse the VLD/IDCT ones, and try different strategies and hw-dependent specializations for the IDCT routine (AAN algorithm, packed FP16, cooperative matrices).

Pretty much reverse engineered: https://mk.pars.ee/notes/a9ihgynpvdo6003w

NVENC/NVDEC could do part of the processing on the shader cores instead of the fixed-function hardware.